The Engineering Behind Self-Driving Cars

Thesis: Redundant, physically diverse sensing is not an expensive luxury. Camera-only autonomy is an engineering position that trades safety margin for cost, and the record so far favors redundancy.

Two philosophies of autonomous driving are being tested in public simultaneously. One holds that because humans drive with two eyes, cameras and sufficiently good software should suffice. The other holds that a safety-critical system should never depend on a single sensing modality. I found this disagreement fascinating because it is not really about artificial intelligence. It is a classic reliability engineering argument, the same one that governs aircraft and medical devices, playing out on public roads with ordinary people as the surrounding population.

The core technology

Cameras provide high-resolution color and texture but infer depth indirectly and degrade badly in glare, fog, and darkness. LiDAR emits laser pulses and measures return time, producing direct, accurate three-dimensional geometry independent of ambient light. Radar measures range and, through the Doppler shift, velocity directly, and it penetrates rain and dust. Each fails differently, which is the entire point: fusing them means a single environmental condition is unlikely to blind the vehicle. Above the sensors sit perception, prediction, and planning software, and beneath them sit drive-by-wire actuators with redundant power and communication paths, because a steering command must arrive even after a fault. Time synchronization across these subsystems matters enormously, because fusing measurements taken milliseconds apart at highway speed introduces meters of positional error.

Figure 1. The roof-mounted LiDAR, camera, and radar sensor pod on a Waymo vehicle — a physical example of the redundant, multi-modal sensing this essay argues for.

Figure 2. LiDAR, camera, and radar compared across driving tasks. Each sensor is strong where the others are weak, which is why fusing all three performs best across the board.

Recent developments and real problems

Commercial robotaxi services now operate driverless in several cities, and they rely on full sensor suites plus detailed prior maps. LiDAR unit costs have fallen, weakening the main economic objection to including it. Four-dimensional imaging radar adds elevation to traditional radar output. The unresolved problems are the long tail: unusual objects, ambiguous human gestures, construction zones, and emergency vehicles. Safety validation is itself unsolved, since demonstrating a rate of rare failures lower than human drivers would require billions of test kilometers, forcing reliance on simulation whose fidelity is difficult to prove. Weather is a further practical limit, with heavy rain and snow degrading both optical and laser sensing in ways that are hard for the software stack to detect from the inside.

Figure 3. Waymo's driverless operating areas as of 2025 — coverage is still geofenced city by city rather than universal, reflecting how far validation and regulation lag behind the technology.

Outlook and solutions

I expect geofenced, sensor-rich autonomy to keep expanding city by city while universal self-driving remains distant. The right responses are structural: mandatory public reporting of disengagements and collisions in comparable formats, standardized scenario libraries for validation, functional safety architectures with independent monitoring channels that can execute a minimal risk maneuvers if the main stack becomes uncertain, and vehicle-to-infrastructure signals at intersections where blocking is unavoidable. Regulation should specify measurable safety outcomes rather than prescribe sensors, but transparency must be non-negotiable. Insurance markets may end up enforcing this more effectively than regulators, because premiums price residual risk whether or not a manufacturer chooses to disclose it.

Conclusion

My position is that removing a sensor to save money is a legitimate engineering choice only if the residual risk is quantified, and that has not consistently happened. The insight I take from this topic is about professional ethics as much as technology: engineers working on safety-critical systems inherit an obligation to argue for margin even when margin is unpopular. Redundancy is what expertise looks like when nothing goes wrong.

Sources

1. IIHS (2026). Waymo's driverless cars crash less often than people — independent analysis of ~50 million driverless miles. https://www.iihs.org/news/detail/waymos-driverless-cars-crash-less-often-than-people

2. K. D. Kusano et al. (2025). Comparison of Waymo Rider-Only Crash Rates by Crash Type to Human Benchmarks at 56.7 Million Miles. Traffic Injury Prevention. https://www.tandfonline.com/doi/full/10.1080/15389588.2025.2499887